workers = 1 can now be overridden by
specifying workers = I(1).Some warnings and errors showed the wrong call.
print() for FutureResult would report
captured conditions all with class list, instead of their
condition classes.
R CMD check --as-cran on R-devel and MS Windows
would trigger a NOTE on “Check: for detritus in the temp directory” and
“Found the following files/directories: ‘Rscript1349cb8aeeba0’ …”. There
were two package tests that explicitly created PSOCK cluster without
stopping them. A third test launched multisession future without
resolving it, which prevented the PSOCK worker to terminate. This was
not detected in R 4.2.0. It is not a problem on macOS and Linux, because
there background workers are automatically terminated when the main R
session terminates.R options and environment variables are now reset on the workers after future is resolved as they were after any packages required by the future has been loaded and attached. Previously, they were reset to what they were before these were loaded and attached. In addition, only pre-existing R options and environment variables are reset. Any new ones added are not removed for now, because we do not know which added R options or environment variables might have been added from loading a package and that are essential for that package to work.
If it was changed while evaluating the future expression, the current working directory is now reset when the future has been resolved.
futureSessionInfo() gained argument
anonymize. If TRUE (default), host and user names are
anonymized.
futureSessionInfo() now also report on the main R
session details.
The bug fix in future 1.22.0 that addressed the
issue where object a in
future(fcn(), globals = list(a = 42, fcn = function() a))
would not be found has been redesigned in a more robust way.
Use of packages such as data.table and ff in cluster and multisession futures broke in future 1.25.0. For data.table, we saw “Error in setalloccol(ans) : verbose must be TRUE or FALSE”. For ff, we saw “Error in splitted$path[nopath] <- getOption(”fftempdir”) : replacement has length zero”. See ‘Significant Changes’ for why and how this was fixed.
The deprecation warning for using local = FALSE was
silenced for sequential futures since future
1.25.0.
futureCall() ignored arguments stdout,
conditions, earlySignal, label,
and gc.
Strategy ‘transparent’ was deprecated in future
1.24.0 and is now defunct. Use
plan(sequential, split = TRUE) instead.
Strategy ‘multiprocess’ was deprecated in future
1.20.0, and ‘remote’ was deprecated in future 1.24.0.
Since then, attempts to use them in plan() would produce a
deprecation warning, which was limited to one per R session. Starting
with this release, this warning is now produced whenever using
plan() with these deprecated future strategies.
Now
f <- future(..., stdout = structure(TRUE, drop = TRUE))
will cause the captured standard output to be dropped from the future
object as soon as it has been relayed once, for instance, by
value(f). Similarly,
conditions = structure("conditions", drop = TRUE) will
drop captured non-error conditions as soon as they have been relayed.
This can help decrease the amount of memory used, especially if there
are many active futures.
Now resolve() respects option
future.wait.interval. Previously, it was hardcoded to poll
for results every 0.1 seconds.
value() will only attempt to recover UTF-8 symbols
in the captured standard output if the future was evaluated on an MS
Windows that does not support capturing of UTF-8 symbols. Support for
UTF-8 capturing on also MS Windows was added in R 4.2.0, but it
typically requires an up-to-date MS Windows 10 or MS Windows Server
2022.future.wait.interval was
decreased from 0.2 seconds to 0.01 seconds. This controls the polling
frequency for finding an available worker when all workers are currently
busy. Starting with this release, this option also controls the polling
frequency of resolve().plan(multicore, workers = 2) and
plan(sequential, split = TRUE) introduced breaking side
effects to the futures evaluated.future(..., seed = TRUE) forwards the RNG state in
the calling R session. Previously, it would leave it intact.plan() and tweak() preserve calls in
arguments,
e.g. plan(multisession, workers = 2, rscript_startup = quote(options(socketOptions="no-delay"))),
and tweak(..., abc = quote(x == y)).nbrOfFreeWorkers() would produce “Error:
‘is.character(name)’ is not TRUE” for
plan(multisession, workers = 1).
Internal calls to
FutureRegistry(action = "collect-first") and
FutureRegistry(action = "collect-last") could signal errors
early when polling resolved().
Strategy ‘remote’ is deprecated in favor of ‘cluster’. The
plan() function will give an informative deprecation
warning when ‘remote’ is used. For now, this warning is given only once
per R session.
Strategy ‘transparent’ is deprecated in favor of ‘sequential’
with argument split = TRUE set. The plan()
function will give an informative deprecation warning when ‘transparent’
is used. For now, this warning is given only once per R
session.
plan() now produces a one-time warning if a
‘transparent’ strategy is set. The warning reminds the user that
‘transparent’ should only be used for troubleshooting purposes and never
be used in production. These days
plan(sequential, split = TRUE) together with
debug() is probably a better approach for troubleshooting.
The long-term plan is to deprecate the ‘transparent’ strategy.
Support for persistent = TRUE with multisession
futures is defunct.
\u2713) would be
relayed as <U+2713> (8 characters). The reason for
this is a limitation in R itself on MS Windows. Now,
value() attempts to recover such MS Windows output to UTF-8
before relaying it. There is an option for disabling this new
feature.quit() must not be used in
forked R processes.future(..., seed) will set the random seed as late
as possible just before the future expression is evaluated. Previously
it was done before package dependencies where attached, which could lead
to non-reproduce random numbers in case a package dependency would
update the RNG seed when attached.values(), which has been deprecated since
future 1.20.0, is now defunct. Use value()
instead.
Support for persistent = TRUE with multisession
futures is defunct. If still needed, a temporary workaround is to use
cluster futures. However, it is likely that support for
persistent will eventually be deprecated for all future
backends.
Argument value of resolve(), deprecated
since future 1.15.0, is defunct in favor of argument
result.
parallel::makeCluster(..., type = "FORK"). This test is
disabled on macOS, where it appears that the main R session becomes
unstable after the FORK node is terminated.A lazy future remains a generic future until it is launched, which means it is not assigned a future backend class until launched.
Argument seed for futureAssign() and
futureCall() now defaults to FALSE just like for
future().
R_FUTURE_* environment variables are now only read
when the future package is loaded, where they set the
corresponding future.* option. Previously, some of these
environment variables were queried by different functions as a fallback
to when an option was not set. By only parsing them when the package is
loaded, it decrease the overhead in functions, and it clarifies that
options can be changed at runtime whereas environment variables should
only be set at startup.
The overhead of initiating futures have been significantly
reduced. For example, the roundtrip time for
value(future(NULL)) is about twice as fast for
‘sequential’, ‘cluster’, and ‘multisession’ futures. For ‘multicore’
futures the roundtrip speedup is about 20%. The speedup comes from
pre-compiling the R expression that will be used to resolve the future
expression into R expression templates which then can quickly compiled
for each future. This speeds up the creation of these expression by ~10
times, compared when re-compiling them each time.
The default timeout for resolved() was decreased
from 0.20 to 0.01 seconds for cluster/multisession and multicore
futures, which means they will spend less time waiting for results when
they are not available.
Analogously to how globals may be scanned for “non-exportable”
objects when option future.globals.onReference is set to
"error" or "warning", value()
will now check for similar problems in the resolved value object. An
example of this is
f <- future(xml2::read_xml("<body></body>")),
which will result in an invalid xml_document object if run
in parallel, because such objects cannot be transferred between R
processes.
In addition to specify which condition classes to be captured and
relayed, it is now possible to also specify condition classes to be
ignored. For example,
conditions = structure("condition", exclude = "message")
captures all conditions but message conditions.
Now cluster futures use homogeneous = NULL as the
default instead of homogeneous = TRUE. The new default will
result in the parallelly package trying to infer
whether TRUE or FALSE should be used based on the workers
argument.
Now the the post-mortem analysis report of multicore and cluster futures in case their results could not be retrieved include information on globals and their sizes, and if some of them are non-exportable. A similar, detailed report is also produced when a cluster future fails to set up and launch itself on a parallel worker.
if option future.fork.multithreading.enable is
FALSE, RcppParallel, in addition to
OpenMP, is forced to run with a single threaded
whenever running in a forked process (=‘multicore’ futures). This is
done by setting environment variable
RCPP_PARALLEL_NUM_THREADS to 1.
Add futureSessionInfo() to get a quick overview of
the future framework, its current setup, and to run simple tests on
it.
Now plan(multicore) warns immediately if multicore
processing, that is, forked processing, is not supported, e.g. when
running in the RStudio Console.
plan(multiprocess, workers = n) did not warn about
‘multiprocess’ being deprecated when argument workers was
specified.
getGlobalsAndPackages() could throw a false error on
“Did you mean to create the future within a function? Invalid future
expression tries to use global ... variables that do not
exist: ... is solely part of a
formula or used in some S4 generic functions.
When enabled, option future.globals.onReference
could falsely alert on ‘Detected a non-exportable reference
(externalptr) in one of the globals (<unknown>) used in the future
expression’ in globals, e.g. when using future.apply or
furrr map-reduce functions when using a ‘multisession’
backend.
future(fcn(), globals = list(a = 42, fcn = function() a))
would fail with “Error in fcn() : object ‘a’ not found” when using
sequential or multicore futures. This affected also map-reduce calls
such as
future.apply::future_lapply(1, function(x) a, future.globals = list(a = 42)).
Resolving a ‘sequential’ future without globals would result in
internal several ...future.* objects being written to the
calling environment, which might be the global environment.
Environment variable R_FUTURE_PLAN would propagate
down with nested futures, forcing itself onto also nested future plans.
Now it is unset in nested futures, resulting in a sequential future
strategy unless another was explicitly set by
plan().
Transparent futures no longer warn about
local = FALSE being deprecated. Although
local = FALSE is being deprecated, it is still used
internally by ‘transparent’ futures for a while longer. Please do not
use ‘transparent’ futures in production code and never in a
package.
remote() could produce an error on “object
‘homogeneous’ not found”.
nbrOfFreeWorkers() for ‘cluster’ futures assumed
that the current plan is set to cluster too.
In order to handle them conditionally higher up in the call
chain, warnings and errors produced from using the random number
generator (RNG) in a future without declaring the intention to use one
are now of class RngFutureWarning and
RngFutureError, respectively. Both of these classes
inherits from RngFutureCondition.
Now run-time errors from resolving a future take precedence over
RngFutureError:s. That is,
future({ rnorm(1); log("a") }, seed = FALSE) will signal an
error ‘log(“a”)’ instead of an RNG error when option
future.rng.onMisuse is set to
"error".
nbrOfFreeWorkers() to query how many workers are
free to take on futures immediately. Until all third-party future
backends have implemented this, some backends might produce an error
saying it is not yet supported.future(..., seed = TRUE) with ‘sequential’ futures
would set the RNG kind of the parent process. Now it behaves the same
regardless of future backend.
Signaling immediateCondition:s with ‘multicore’
could result in “Error in save_rds(obj, file) : save_rds() failed to
rename temporary save file
‘/tmp/RtmpxNyIyK/progression21f3f31eadc.rds.tmp’ (NA bytes; last
modified on NA) to ‘/tmp/RtmpxNyIyK/progression21f3f31eadc.rds’ (NA
bytes; last modified on NA)”. There was an assertion at the end of the
internal save_rds() function that incorrectly assumed that
the target file should exist. However, the file might have already been
processed and removed by the future in the main R session.
value() with both a run-time error and an RNG
mistake would signal the RNG warning instead of the run-time error when
the for-internal-use-only argument signal was set to
FALSE.
Due to a mistake introduced in future 1.20.0,
the package would end up assigning a .packageVersion object
to the global environment when loaded.
future::plan("multisession") would produce ‘Error in if
(debug) mdebug(“covr::package_coverage() workaround …”) : argument is
not interpretable as logical’ if and only if the covr
package was loaded.Strategy ‘multiprocess’ is deprecated in favor of either
‘multisession’ or ‘multicore’, depending on operating system and R
setup. The plan() function will give an informative
deprecation warning when ‘multiprocess’ is used. This warning is given
only once per R session.
Launching R or Rscript with command-line option
--parallel=n, where n > 1, will now use ‘multisession’
as future strategy. Previously, it would use ‘multiprocess’, which is
now deprecated.
Support for local = FALSE is deprecated. For the
time being, it remains supported for ‘transparent’ futures and ‘cluster’
futures that use persistent = TRUE. However, note that
persistent = TRUE will also deprecated at some point in the
future. These deprecations are required in order to further standardize
the Future API across various types of parallel backends.
Now multisession workers inherit the package library path from
the main R session when they are created, that is, when calling
plan(multisession). To avoid this, use
plan(multisession, rscript_libs = NULL), which is an
argument passed down to makeClusterPSOCK(). With this
update, ‘sequential’, ‘multisession’, and ‘multicore’ futures see the
exact same library path.
Several functions for managing parallel-style
processing have been moved to a new parallelly package.
Specifically, functions availableCores(),
availableWorkers(), supportsMulticore(),
as.cluster(), autoStopCluster(),
makeClusterMPI(), makeClusterPSOCK(), and
makeNodePSOCK() have been moved. None of them are specific
to futures per se and are likely useful elsewhere too. Also, having them
in a separate, standalone package will speed up the process of releasing
any updates to these functions. The code base of the
future package shrunk about 10-15% from this migration.
For backward compatibility, the migrated functions remain in this
package as re-exports.
Setting up a future strategy with argument
split = TRUE will cause the standard output and non-error
conditions to be split (“tee:d”) on the worker’s end, while still
relaying back to the main R session as before. This can be useful when
debugging with browse() or debug(), e.g.
plan(sequential, split = TRUE). Without it, debug output is
not displayed.
Now multicore futures relay immediateCondition:s in
a near-live fashion.
It is now possible to pass any arguments that
makeClusterPSOCK() accepts in the call to
plan(cluster, ...) and
plan(multisession, ...). For instance, to set the working
directory of the cluster workers to a temporary folder, pass argument
rscript_startup = "setwd(tempdir())". Another example is
rscript_libs = c(libs, "*") to prepend the library path on
the worker with the paths in libs.
plan() and tweak() check for even more
arguments that must not be set by either of them. Specifically, attempts
to adjust the following arguments of future() will result
in an error: conditions, envir,
globals, packages, stdout, and
substitute in addition to already validated
lazy and seed.
tweak() now returns a wrapper function that calls
the original future strategy function with the modified defaults.
Previously, it would make a copy of the original function with modified
argument defaults. This new approach will make it possible to introduce
new future arguments that can be modified by tweak() and
plan() without having to update every future backend
package, e.g. the new split = TRUE argument.
Add a ‘Best Practices for Package Developers’ vignette.
Add a ‘How the Future Framework is Validated’ vignette.
substitute = TRUE.Since last version, future 1.19.1,
future(..., conditions = character(0L)) would no longer
avoid intercepting conditions as intended; instead, it muffles all
conditions. From now on, use conditions = NULL.
Relaying of immediateCondition:s was not near-live
for multisession and cluster if the underlying PSOCK cluster used
useXDR=FALSE for communication.
print() for Future would also print any attributes
of its environment.
The error message produced by nbrOfWorkers() was
incomplete.
Renamed environment variable
R_FUTURE_MAKENODEPSOCK_tries used by
makeClusterPSOCK() to
R_FUTURE_MAKENODEPSOCK_TRIES.
The Mandelbrot demo would produce random numbers without declaring so.
Strategy ‘multiprocess’ is deprecated in favor of either ‘multisession’ or ‘multicore’, depending on operating system and R setup.
values() is deprecated. Use value()
instead.
All backward compatible code for the legacy, defunct, internal
Future element value is now removed. Using or
relying on it is an error.
... as a globals, rather than via
arguments, in higher-level map-reduce APIs such as
future.apply and furrr, arguments in
... could produce an error on “unused argument”.Futures detect when random number generation (RNG) was used to
resolve them. If a future uses RNG without parallel RNG was requested,
then an informative warning is produced. To request parallel RNG,
specify argument seed, e.g.
f <- future(rnorm(3), seed = TRUE) or
y %<-% { rnorm(3) } %seed% TRUE. Higher-level map-reduce
APIs provide similarly named “seed” arguments to achieve the same. To,
escalate these warning to errors, set option
future.rng.onMisuse to "error". To silence
them, set it to "ignore".
Now, all non-captured conditions are muffled, if possible. For
instance,
future(warning("boom"), conditions = c("message")) will
truly muffle the warning regardless of backend used. This was needed to
fix below bug.
makeClusterPSOCK() will now retry to create a
cluster node up to tries (default: 3) times before giving
up. If argument port species more than one port
(e.g. port = "random") then it will also attempt find a
valid random port up to tries times before giving up. The
pre-validation of the random port is only supported in R (>= 4.0.0)
and skipped otherwise.
makeClusterPSOCK() skips shell quoting of the
elements in rscript if it inherits from
AsIs.
makeClusterPSOCK(), or actually
makeNodePSOCK(), gained argument quiet, which
can be used to silence output produced by
manual = TRUE.
If multithreading is disabled but multicore futures fail to
acknowledge the setting on the current system, then an informative
FutureWarning is produced by such futures.
Now availableCores() better supports Slurm.
Specifically, if environment variable SLURM_CPUS_PER_TASK
is not set, which requires that option
--slurm-cpus-per-task=n is specified and
SLURM_JOB_NUM_NODES=1, then it falls back to using
SLURM_CPUS_ON_NODE, e.g. when using
--ntasks=n.
Now availableCores() and
availableWorkers() supports LSF/OpenLava. Specifically,
they acknowledge environment variable LSB_DJOB_NUMPROC and
LSB_HOSTS, respectively.
plan(multisession),
plan(cluster, workers = <number>), and
makeClusterPSOCK() which they both use internally, sets up
localhost workers twice as fast compared to versions since
future 1.12.0, which brings it back to par with a
bare-bone
parallel::makeCluster(..., setup_strategy = "sequential")
setup. The slowdown was introduced in future 1.12.0
(2019-03-07) when protection against leaving stray R processes behind
from failed worker startup was implemented. This protection now makes
use of memoization for speedup.Sequential and multicore backends, but not multisession, would
produce errors on “‘…’ used in an incorrect context” in cases where
... was part of argument globals and not the
evaluation environment.
Contrary to other future backends, any conditions produced while
resolving a sequential future using
future(..., conditions = character()) would be signaled,
although the most reasonable expectation would be that they are
silenced. Now, all non-captured conditions are muffled, if
possible.
Option future.rng.onMisuse was not passed down to
nested futures.
Disabling multithreading in forked processes by setting R option
future.fork.multithreading.enable or environment variable
R_FUTURE_FORK_MULTITHREADING_ENABLE to FALSE
would cause multicore futures to always return value 1L.
This bug was introduced in future 1.17.0
(2020-04-17).
getGlobalsAndPackages() did not always return a
globals element that was of class
FutureGlobals.
getGlobalsAndPackages(..., globals) would
recalculate total_size even when it was already calculated
or known to be zero.
getGlobalsAndPackages(Formula::Formula(~ x)) would
produce “the condition has length > 1” warnings (which will become
errors in future R versions).
persistent = TRUE with multisession futures
is deprecated.print() on RichSOCKcluster gives
information not only on the name of the host but also on the version of
R and the platform of each node (“worker”), e.g. “Socket cluster with 3
nodes where 2 nodes are on host ‘localhost’ (R version 4.0.0
(2020-04-24), platform x86_64-w64-mingw32), 1 node is on host ‘n3’ (R
version 3.6.3 (2020-02-29), platform x86_64-pc-linux-gnu)”.
Error messages from cluster future failures are now more
informative than “Unexpected result (of class ‘NULL’ !=
‘FutureResult’)”. For example, if the future package is
not installed on the worker, then the error message clearly says so.
Even, if there is an unexpected result error from a PSOCK cluster
future, then the error produced give extra information on node where it
failed, e.g. “Unexpected result (of class ‘NULL’ != ‘FutureResult’)
retrieved for ClusterFuture future (label = ‘ClusterFuture
worker (‘RichSOCKnode’ #1 on host ‘n3’ (R version 3.6.3 (2020-02-29),
platform x86_64-pc-linux-gnu)) is out of sync.”
It is now possible to set environment variables on workers before
they are launched by makeClusterPSOCK() by specify them as
as "<name>=<value>" as part of the
rscript vector argument,
e.g. rscript = c("ABC=123", "DEF='hello world'", "Rscript").
This works because elements in rscript that match regular
expression [[:alpha:]_][[:alnum:]_]*=.* are no longer shell
quoted.
makeClusterPSOCK() now returns a cluster that in
addition to inheriting from SOCKcluster it will also
inherit from RichSOCKcluster.
Made makeClusterPSOCK() and
makeNodePSOCK() agile to the name change from
parallel:::.slaveRSOCK() to
parallel:::.workRSOCK() in R (>= 4.1.0).
makeClusterPSOCK(..., rscript) will not try to
locate rscript[1] if argument homogeneous is
FALSE (or inferred to be FALSE).
makeClusterPSOCK(..., rscript_envs) would result in
a syntax error when starting the workers due to non-ASCII quotation
marks if option useFancyQuotes was not set to
FALSE.
plan(list(...)) would produce ‘Error in
UseMethod(“tweak”) : no applicable method for ’tweak’ applied to an
object of class “list”’ if a non-function object named ‘list’ was on the
search path.
plan(x$abc) with x <- list(abc = sequential)
would produce ‘Error in UseMethod(“tweak”) : no applicable method for
’tweak’ applied to an object of class “c(‘FutureStrategyList’,
‘list’)”’.
TESTS: R_FUTURE_FORK_ENABLE=false R CMD check ...
would produce ‘Error: connections left open: …’ when checking the
‘multiprocess’ example.
Support for persistent = TRUE with multisession
futures is deprecated. If still needed, a temporary workaround is to use
cluster futures. However, it is likely that support for
persistent will eventually be deprecated for all future
backends.
Options future.globals.method,
future.globals.onMissing, and
future.globals.resolve are deprecated and produce warnings
if set. They may only be used for troubleshooting purposes because they
may affect how futures are evaluated, which means that reproducibility
cannot be guaranteed elsewhere.
values() to value() to clean up
and simplify the API.makeClusterPSOCK() gained argument
rscript_envs for setting environment variables in workers
on startup,
e.g. rscript_envs = c(FOO = "3.14", "BAR").
Now the result of a future holds session details in case an error occurred while evaluating the future.
_R_CHECK_LIMIT_CORES_ set. To
better emulate CRAN submission checks, the future
package will, when loaded, set this environment variable to ‘TRUE’ if
unset and if R CMD check is running. Note that
future::availableCores() respects
_R_CHECK_LIMIT_CORES_ and returns at most 2L
(two cores) if detected.Any globals named version and
has_future would be overwritten with “garbage” values
internally.
Disabling of multi-threading when using ‘multicore’ futures did not work on all platforms.
values() S3 methods have been renamed to
value() since they are closely related to the original
purpose value(). The values() methods will
continue to work but will soon be formally deprecated and later be made
defunct and finally be removed. Please replace all values()
with value() calls.oplan <- plan(new_strategy) returns the list of
all nested strategies previously set, instead of just the strategy on
top of this stack. This makes it easier to temporarily use another plan.
For the old behavior, use
oplan <- plan(new_strategy)[[1]].Now value() detects if a
future(..., seed = FALSE) call generated random numbers,
which then might give unreliable results because non-parallel safe,
non-statistically sound random number generation (RNG) was used. If
option future.rng.onMisuse is "warning", a
warning is produced. If "error", an error is produced. If
"ignore" (default), the mistake is silently ignored. Using
seed = NULL is like seed = FALSE but without
performing the RNG validation.
For convenience, argument seed of
future() may now also be an ordinary single integer random
seed. If so, a L’Ecuyer-CMRG RNG seed is created from this seed. If
seed = TRUE, then a L’Ecuyer-CMRG RNG seed based on the
current RNG state is used. Use seed = FALSE when it is
known that the future does not use RNG.
ClusterFuture:s now relay
immediateCondition:s back to the main process momentarily
after they are signaled and before the future is resolved.
future.fork.multithreading.enable or environment variable
R_FUTURE_FORK_MULTITHREADING_ENABLE to FALSE.
This requires that RhpcBLASctl package is installed.
Parallelization via multi-threaded processing (done in native code by
some packages and externally library) while at the same time using
forked (aka “multicore”) parallel processing is unstable in some cases.
Note that this is not only true when using plan(multicore)
but also when using, for instance, parallel::mclapply().
This is in beta so the above names and options might change later.Evaluation of futures could fail if the global environment
contained functions with the same names as a small set of base
R functions, e.g. raw(), list(), and
options().
future(alist(a =)) would produce “Error in
objectSize_list(x, depth = depth - 1L) : argument”x_kk” is missing, with
no default”
Future and FutureResult objects with an
internal version 1.7 or older have been deprecated since 1.14.0 (July
2019) and are now defunct.
Defunct hidden argument progress of
resolve(), and hidden arguments/fields
condition and calls of
FutureResult are now gone.
makeClusterPSOCK()
draws a random port from (when argument port is not
specified) can now be controlled by environment variable
R_FUTURE_RANDOM_PORTS. The default range is still
11000:11999 as with the parallel
package.resolved() in
future 1.15.0 would cause lazy futures to block if all
workers were occupied.resolved() will now launch lazy futures.Now the “visibility” of future values is recorded and reflected
by value().
Now option future.globals.onReference defaults to
environment variable R_FUTURE_GLOBALS_ONREFERENCE.
?makeClusterPSOCK
with instructions on how to troubleshoot when the setup of local and
remote clusters fail.values() would resignal
immediateCondition:s despite those should only be signaled
at most once per future.
makeClusterPSOCK() could produce warnings like
“cannot open file
‘/tmp/alice/Rtmpi69yYF/future.parent=2622.a3e32bc6af7.pid’: No such
file”, e.g. when launching R workers running in Docker
containers.
Package would set or update the RNG state of R
(.Random.seed) when loaded, which could affect RNG
reproducibility.
Package could set .Random.seed to NULL, instead of
removing it, which in turn would produce a warning on “‘.Random.seed’ is
not an integer vector but of type ‘NULL’, so ignored” when the next
random number generated.
Now a future assignment to list environments produce more informative error messages if attempting to assign to more than one element.
makeClusterMPI() did not work for MPI clusters with
comm other than 1.
Argument value of resolve() is
deprecated. Use result instead.
Use of internal argument evaluator to
future() is now defunct.
All types of conditions are now captured and relayed. Previously,
only conditions of class message and warning
were relayed.
If one of the futures in a collection produces an error, then
values() will signal that error as soon as it is detected.
This means that while calling values() guarantees to
resolve all futures, it does not guarantee that the result from all
futures are gathered back to the master R session before the error is
relayed.
values() now relays stdout and signal
as soon as possible as long as the standard output and the conditions
are relayed in their original order.
If a captured condition can be “muffled”, then it will be
muffled. This helps to prevent conditions from being handled twice by
condition handlers when futures are evaluated in the main R session,
e.g. plan(sequential). Messages and warnings were already
muffled in the past.
Forked processing is considered unstable when running R from
certain environments, such as the RStudio environment. Because of this,
‘multicore’ futures have been disabled in those cases since
future 1.13.0. This change caught several RStudio users
by surprise. Starting with future 1.14.0, an
informative one-time-per-session warning will be produced when attempts
to use ‘multicore’ is made in non-supported environments such as
RStudio. This warning will also be produced when using ‘multiprocess’,
which will fall back to using ‘multisession’ futures. The warning can be
disabled by setting R option
future.supportsMulticore.unstable, or environment variable
FUTURE_SUPPORTSMULTICORE_UNSTABLE to
"quiet".
Now option future.startup.script falls back to
environment variable R_FUTURE_STARTUP_SCRIPT.
Conditions inheriting immediateCondition are
signaled as soon as possible. Contrary to other types of conditions,
these will be signaled only once per future, despite being
collected.
Early signaling did not take place for resolved()
for ClusterFuture and
MulticoreFuture.
When early signaling was enabled, functions such as
resolved() and resolve() would relay captured
conditions multiple times. This would, for instance, result in the same
messages and warnings being outputted more than once. Now it is only
value() that will resignal conditions.
The validation of connections failed to detect when the
connection had been serialized (= a NIL external pointer)
on some macOS systems.
Argument progress of resolve() is now
defunct (was deprecated since future 1.12.0). Option
future.progress is ignored. This will make room for other
progress-update mechanisms that are in the works.
Usage of internal argument evaluator to
future() is now deprecated.
Removed defunct argument output from
FutureError().
FutureResult fields/arguments condition
and calls are now defunct. Use conditions
instead.
Future and FutureResult objects with an
internal version 1.7 or older are deprecated and will eventually become
defunct. Future backends that implement their own Future
classes should update to implement a result() method
instead of a value() method for their Future
classes. All future backends available on CRAN and Bioconductor have
already been updated accordingly.
help("supportsMulticore") for more details, e.g. how to
re-enable process forking. Note that parallelization via ‘multisession’
is unaffected and will still work as before. Also, when forked
processing is disabled, or otherwise not supported, using
plan("multiprocess") will fall back to using ‘multisession’
futures.Forked processing can be disabled by setting R option
future.fork.enable to FALSE (or environment variable
R_FUTURE_FORK_ENABLE=false). When disabled, ‘multicore’
futures fall back to a ‘sequential’ futures even if the operating system
supports process forking. If set of TRUE, ‘multicore’ will not fall back
to ‘sequential’. If NA, or not set (the default), a set of
best-practices rules will decide whether forking is enabled or not. See
help("supportsMulticore") for more details.
Now availableCores() also recognizes PBS environment
variable NCPUS, because the PBSPro scheduler does not set
PBS_NUM_PPN.
If, option future.availableCores.custom is set to a
function, then availableCores() will call that function and
interpret its value as number of cores. Analogously, option
future.availableWorkers.custom can be used to specify a
hostnames of a set of workers that availableWorkers() sees.
These new options provide a mechanism for anyone to customize
availableCores() and availableWorkers() in
case they do not (yet) recognize, say, environment variables that are
specific the user’s compute environment or HPC scheduler.
makeClusterPSOCK() gained support for argument
rscript_startup for evaluating one or more R expressions in
the background R worker prior to the worker event loop launching. This
provides a more convenient approach than having to use, say,
rscript_args = c("-e", sQuote(code)).
makeClusterPSOCK() gained support for argument
rscript_libs to control the R package library search path
on the workers. For example, to prepend the folder
~/R-libs on the workers, use
rscript_libs = c("~/R-libs", "*"), where "*"
will be resolved to the current .libPaths() on the
workers.
Debug messages are now prepended with a timestamp.
makeClusterPSOCK() did not shell quote the Rscript
executable when running its pre-tests checking whether localhost Rscript
processes can be killed by their PIDs or not.value of resolve() has been
renamed to result to better reflect that not only values
are collected when this argument is used. Argument value
still works for backward compatibility, but will eventually be formally
deprecated and then defunct.If makeClusterPSOCK() fails to create one of many
nodes, then it will attempt to stop any nodes that were successfully
created. This lowers the risk for leaving R worker processes
behind.
Future results now hold the timestamps when the evaluation of the future started and finished.
Functions no longer produce “partial match of ‘condition’ to
‘conditions’” warnings with
options(warnPartialMatchDollar = TRUE).
When future infix operators (%conditions%,
%globals%, %label%, %lazy%,
%packages%, %seed%, and %stdout%)
that are intended for future assignments were used in the wrong context,
they would incorrectly be applied to the next future created. Now
they’re discarded.
makeClusterPSOCK() in future (>=
1.11.1) produced warnings when argument rscript had
length(rscript) > 1.
Validation of L’Ecuyer-CMRG RNG seeds failed in recent R devel.
With options(OutDec = ","), the default value of
several argument would resolve to NA_real_ rather than a
numeric value resulting in errors such as “is.finite(alpha) is not
TRUE”.
Argument progress of resolve() is now
deprecated.
Argument output of FutureError() is now
defunct.
FutureError no longer inherits
simpleError.
makeClusterPSOCK() fails to connect to a worker,
it produces an error with detailed information on what could have
happened. In rare cases, another error could be produced when generating
the information on what the workers PID is.The defaults of several arguments of
makeClusterPSOCK() and makeNodePSOCK() can now
be controlled via environment variables in addition to R options that
was supported in the past. An advantage of using environment variables
is that they will be inherited by child processes, also nested
ones.
The printing of future plans is now less verbose when the
workers argument is a complex object such as a PSOCK
cluster object. Previously, the output would include verbose output of
attributes, etc.
R CMD check is running or not. If it is, then a few
future-specific environment variables are adjusted such that the tests
play nice with the testing environment. For instance, it sets the socket
connection timeout for PSOCK cluster workers to 120 seconds (instead of
the default 30 days!). This will lower the risk for more and more zombie
worker processes cluttering up the test machine (e.g. CRAN servers) in
case a worker process is left behind despite the main R processes is
terminated. Note that these adjustments are applied automatically to the
checks of any package that depends on, or imports, the
future package.makeClusterPSOCK() would fail to connect to a
worker, for instance due to a port clash, then it would leave the R
worker process running - also after the main R process terminated. When
the worker is running on the same machine,
makeClusterPSOCK() will now attempt to kill such stray R
processes. Note that parallel::makePSOCKcluster() still has
this problem.The future call stack (“traceback”) is now recorded when the
evaluation of a future produces an error. Use backtrace()
on the future to retrieve it.
Now futureCall() defaults to
args = list() making is easier to call functions that do
not take arguments,
e.g. futureCall(function() 42).
plan() gained argument .skip = FALSE.
When TRUE, setting the same future strategy as already set will be
skipped, e.g. calling plan(multisession) consecutively will
have the same effect as calling it just once.
makeClusterPSOCK() produces more informative error
messages whenever the setup of R workers fails. Also, its verbose
messages are now prefixed with [local output] to help
distinguish the output produced by the current R session from that
produced by background workers.
It is now possible to specify what type of SSH clients
makeClusterPSOCK() automatically searches for and in what
order, e.g.
rshcmd = c("<rstudio-ssh>", "<putty-plink>").
Now makeClusterPSOCK() preserves the global RNG
state (.Random.seed) also when it draws a random port
number.
makeClusterPSOCK() gained argument
rshlogfile.
Cluster futures provide more informative error messages when the communication with the worker node is out of sync.
Argument stdout was forced to TRUE when using
single-core multicore or single-core multisession futures.
When evaluated in a local environment,
futureCall(..., globals = "a") would set the value of
global a to NULL, regardless if it exists or not and what
its true value is.
makeClusterPSOCK(..., rscript = "my_r") would in
some cases fail to find the intended my_r
executable.
ROBUSTNESS: A cluster future, including a multisession one, could retrieve results from the wrong workers if a new set of cluster workers had been set up after the future was created/launched but before the results were retrieved. This could happen because connections in R are indexed solely by integers which are recycled when old connections are closed and new ones are created. Now cluster futures assert that the connections to the workers are valid, and if not, an informative error message is produced.
Calling result() on a non-resolved
UniprocessFuture would signal evaluation errors.
future::future_lapply(). Please use the
one in the future.apply package instead.Add support for manually specifying globals in addition to those
that are automatically identified via argument globals or
%globals%. Two examples are
globals = structure(TRUE, add = list(a = 42L, b = 3.14))
and globals = structure(TRUE, add = c("a", "b")).
Analogously, attribute ignore can be used to exclude
automatically identified globals.
The error reported when failing to retrieve the results of a future evaluated on a localhost cluster/multisession worker or a forked/multicore worker is now more informative. Specifically, it mentions whether the worker process is still alive or not.
Add makeClusterMPI(n) for creating MPI-based
clusters of a similar kind as
parallel::makeCluster(n, type = "MPI") but that also
attempts to workaround issues where parallel::stopCluster()
causes R to stall.
makeClusterPSOCK() and makeClusterMPI()
gained argument autoStop for controlling whether the
cluster should be automatically stopped when garbage collected or
not.
BETA: Now resolved() for ClusterFuture
is non-blocking also for clusters of type MPIcluster as
created by
parallel::makeCluster(..., type = "MPI").
plan(multiprocess) would not initiate the
workers. Instead workers would be set up only when the first future was
created.value() is called. This new behavior can be controlled by
the argument stdout to future() or by
specifying the %stdout% operator if a future assignment is
used.R option width is passed down so that standard
output is captured consistently across workers and consistently with the
master process.
Now more future.* options are passed down so that
they are also acknowledged when using nested futures.
Add vignette on ‘Outputting Text’.
CLEANUP: Only the core parts of the API are now listed in the
help index. This was done to clarify the Future API. Help for non-core
parts are still via cross references in the indexed API as well via
help().
When using forced, nested ‘multicore’ parallel processing, such
as,
plan(list(tweak(multicore, workers = 2), tweak(multicore, workers = 2))),
then the child process would attempt to resolve futures owned by the
parent process resulting in an error (on ‘bad error message’).
When using plan(multicore), if a forked worker would
terminate unexpectedly, it could corrupt the master R session such that
any further attempts of using forked workers would fail. A forked worker
could be terminated this way if the user pressed Ctrl-C (the
worker receives a SIGINT signal).
makeClusterPSOCK() produced a warning when
environment variable R_PARALLEL_PORT was set to
random (e.g. as on CRAN).
Printing a plan() could produce an error when the
deparsed call used to set up the plan() was longer than 60
characters.
future::future_lapply() is defunct (gives an error
if called). Please use the one in the future.apply
package instead.
Argument output of FutureError() is
formally deprecated.
Removed all FutureEvaluationCondition classes and
related methods.
getGlobalsAndPackages() gained argument
maxSize.
makeClusterPSOCK() now produces a more informative
warning if environment variable R_PARALLEL_PORT specifies a
non-numeric port.
Now plan() gives a more informative error message in
case it fails, e.g. when the internal future validation fails and
why.
Added UnexpectedFutureResultError to be used by
backends for signaling in a standard way that an unexpected result was
retrieved from a worker.
When the communication between an asynchronous future and a background R process failed, further querying of the future state/results could end up in an infinite waiting loop. Now the failed communication error is recorded and re-signaled if any further querying attempts.
Internal, seldom used myExternalIP() failed to
recognize IPv4 answers from some of the lookup servers. This could in
turn produce another error.
In R (>= 3.5.0), multicore futures would produce multiple warnings originating from querying whether background processes have completed or not. These warnings are now suppressed.
More errors related to orchestration of futures are of class
FutureError to make it easier to distinguish them from
future evaluation errors.
Add support for a richer set of results returned by resolved
futures. Previously only the value of the future expression, which could
be a captured error to be resignaled, was expected. Now a
FutureResult object may be returned instead. Although not
supported in this release, this update opens up for reporting on
additional information from the evaluation of futures, e.g. captured
output, timing and memory benchmarks, etc. Before that can take place,
existing future backend packages will have to be updated
accordingly.
backtrace() returns only the last call that produced
the error. It is unfortunately not possible to capture the call stack
that led up to the error when evaluating a future expression.
value() for MulticoreFuture would not
produce an error when a (forked) background R workers would terminate
before the future expression is resolved. This was a limitation
inherited from the parallel package. Now an informative
FutureError message is produced.
value() for MulticoreFuture would not
signal errors unless they inherited from simpleError - now
it’s enough for them to inherits from error.
value() for ClusterFuture no longer
produces a FutureEvaluationError, but
FutureError, if the connection to the R worker has changed
(which happens if something as drastic as
closeAllConnections() have been called.)
futureCall(..., globals = FALSE) would produce
“Error: second argument must be a list”, because the explicit arguments
where not exported. This could also happen when specifying globals by
name or as a named list.
Nested futures were too conservative in requiring global variables to exist, even when they were false positives.
future::future_lapply() is formally deprecated.
Please use the one in the future.apply package
instead.
Recently introduced FutureEvaluationCondition
classes are deprecated, because they no longer serve a purpose since
future evaluation conditions are now signaled as-is.
future_lapply() has moved to the
future.apply package available on CRAN.Argument workers of future strategies may now also
be a function, which is called without argument when the future strategy
is set up and used as is. For instance,
plan(multiprocess, workers = halfCores) where
halfCores <- function() { max(1, round(availableCores()/ 2)) }
will use half of the number of available cores. This is useful when
using nested future strategies with remote machines.
On Windows, makeClusterPSOCK(), and therefore
plan(multisession) and plan(multiprocess),
will use the SSH client distributed with RStudio as a fallback if
neither ssh nor plink is available on the
system PATH.
Now plan() makes sure that
nbrOfWorkers() will work for the new strategy. This will
help catch mistakes such as plan(cluster, workers = cl)
where cl is a basic R list rather than a
cluster list early on.
Added %packages% to explicitly control packages to
be attached when a future is resolved,
e.g. y %<-% { YT[2] } %packages% "data.table". Note,
this is only needed in cases where the automatic identification of
global and package dependencies is not sufficient.
Added condition classes FutureCondition,
FutureMessage, FutureWarning, and
FutureError representing conditions that occur while a
future is setup, launched, queried, or retrieved. They do not
represent conditions that occur while evaluating the future expression.
For those conditions, new classes
FutureEvaluationCondition,
FutureEvaulationMessage,
FutureEvaluationWarning, and
FutureEvaluationError exists.
if (runif(1) < 1/2) x <- 0; y <- 2 * x.externalptr)
can not be exported, but there are exceptions. By setting options
future.globals.onReference to "warning", a
warning is produced informing the user about potential problems. If
"error", an error is produced. Because there might be false
positive, the default is "ignore", which will cause above
scans to be skipped. If there are non-exportable globals and these tests
are skipped, a run-time error may be produced only when the future
expression is evaluated.The total size of global variables was overestimated, and dramatically so if defined in the global environment and there were are large objects there too. This would sometimes result in a false error saying that the total size is larger than the allowed limit.
An assignment such as x <- x + 1 where the
left-hand side (LHS) x is a global failed to identify
x as a global because the right-hand side (RHS)
x would override it as a local variable. Updates to the
globals package fixed this problem.
makeClusterPSOCK(..., renice = 19) would launch each
PSOCK worker via nice +19 resulting in the error “nice:
‘+19’: No such file or directory”. This bug was inherited from
parallel::makePSOCKcluster(). Now using
nice --adjustment=19 instead.
Protection against passing future objects to other futures did not work for future strategy ‘multicore’.
future_lapply() has moved to the new
future.apply package available on CRAN. The
future::future_lapply() function will soon be deprecated,
then defunct, and eventually be removed from the future
package. Please update your code to make use of
future.apply::future_lapply() instead.
Dropped defunct ‘eager’ and ‘lazy’ futures; use ‘sequential’ instead.
Dropped defunct arguments cluster and
maxCores; use workers instead.
In previous version of the future package the
FutureError class was used to represent both orchestration
errors (now FutureError) and evaluation errors (now
FutureEvaluationError). Any usage of class
FutureError for the latter type of errors is deprecated and
should be updated to FutureEvaluationError.
Now plan() accepts also strings such as
"future::cluster".
Now backtrace(x[[ER]]) works also for
non-environment x:s, e.g. lists.
When measuring the size of globals by scanning their content, for certain types of classes the inferred lengths of these objects were incorrect causing internal subset out-of-range issues.
print() for Future would output one
global per line instead of concatenating the information with
commas.
getGlobalsAndPackages().future_lapply() would give “Error in objectSize.env(x,
depth = depth - 1L): object ‘nnn’ not found” when for instance ‘nnn’ is
part of an unresolved expression that is an argument value.tweak(), and hence plan(), generates a
more informative error message if a non-future function is specified by
mistake, e.g. calling plan(cluster) with the
survival package attached after future
is equivalent to calling plan(survival::cluster) when
plan(future::cluster) was intended.nbrOfWorkers() gave an error with
plan(remote). Fixed by making the ‘remote’ future inherit
cluster (as it should).quit(), but that appeared to have
corrupted the main R session when running on Solaris.Formally defunct ‘eager’ and ‘lazy’ futures; use ‘sequential’ instead.
Dropped previously defunct %<=% and
%=>% operators.
makeClusterPSOCK() now defaults to use the Windows
PuTTY software’s SSH client plink -ssh, if ssh
is not found.
Argument homogeneous of
makeNodePSOCK(), a helper function of
makeClusterPSOCK(), will default to FALSE also if the
hostname is a fully qualified domain name (FQDN), that is, it “contains
periods”. For instance, c('node1', 'node2.server.org') will
use homogeneous = TRUE for the first worker and
homogeneous = FALSE for the second.
makeClusterPSOCK() now asserts that each cluster
node is functioning by retrieving and recording the node’s session
information including the process ID of the corresponding R
process.
Nested futures sets option mc.cores to prevent
spawning of recursive parallel processes by mistake. Because ‘mc.cores’
controls additional processes, it was previously set to zero.
However, since some functions such as mclapply() does not
support that, it is now set to one instead.
makeClusterPSOCK() gained more detailed
descriptions on arguments and what their defaults are.future_lapply() with multicore / multisession
futures, would use a suboptimal workload balancing where it split up the
data in one chunk too many. This is no longer a problem because of how
argument workers is now defined for those type of futures
(see note on top).
future_lapply(), as well as lazy multicore and lazy
sequential futures, did not respect option
future.globals.resolve, but was hardcoded to always resolve
globals (future.globals.resolve = TRUE).
When globals larger than the allowed size (option
future.globals.maxSize) are detected an informative error
message is generated. Previous version introduced a bug causing the
error to produce another error.
Lazy sequential futures would produce an error when resolved if required packages had been detached.
print() would not display globals gathered for lazy
sequential futures.
Added package tests for globals part of formulas part of other
globals, e.g. purrr::map(x, ~ rnorm(.)), which requires
globals (>= 0.10.0).
Now package tests with parallel::makeCluster() not
only test for type = "PSOCK" clusters but also
"FORK" (when supported).
TESTS: Cleaned up test scripts such that the overall processing time for the tests was roughly halved, while preserving the same test coverage.
future_lapply() is now to not
generate RNG seeds (future.seed = FALSE). If proper random
number generation is needed, use future.seed = TRUE. For
more details, see help page.future() and future_lapply() gained
argument packages for explicitly specifying packages to be
attached when the futures are evaluated. Note that the default
throughout the future package is that all globals and
all required packages are automatically identified and gathered, so in
most cases those do not have to be specified manually.
The default values for arguments connectTimeout and
timeout of makeNodePSOCK() can now be
controlled via global options.
Now future_lapply() guarantees that the RNG state of
the calling R process after returning is updated compared to what it was
before and in the exact same way regardless of future.seed
(except FALSE), future.scheduling and future strategy used.
This is done in order to guarantee that an R script calling
future_lapply() multiple times should be numerically
reproducible given the same initial seed.
It is now possible to specify a pre-generated sequence of
.Random.seed seeds to be used for each
FUN(x[[i]], ...) call in
future_lapply(x, FUN, ...).
future_lapply() scans global variables for non-resolved
futures (to resolve them) and calculate their total size once.
Previously, each chunk (a future) would redo this.Now future_lapply(X, FUN, ...) identifies global
objects among X, FUN and ...
recursively until no new globals are found. Previously, only the first
level of globals were scanned. This is mostly thanks to a bug fix in
globals 0.9.0.
A future that used a global object x of a class that
overrides length() would produce an error if
length(x) reports more elements than what can be
subsetted.
nbrOfWorkers() gave an error with
plan(cluster, workers = cl) where cl is a
cluster object created by
parallel::makeCluster(), etc. This prevented for instance
future_lapply() to work with such setups.
plan(cluster, workers = cl) where
cl <- makeCluster(..., type = MPI") would give an
instant error due to an invalid internal assertion.
Previously deprecated arguments maxCores and
cluster are now defunct.
Previously deprecated assignment operators %<=%
and %=>% are now defunct.
availableCores(method = "mc.cores") is now defunct
in favor of "mc.cores+1".
plan(), e.g. plan(cluster) will set up workers
on all cluster nodes. Previously, this only happened when the first
future was created.Renamed ‘eager’ futures to ‘sequential’,
e.g. plan(sequential). The ‘eager’ futures will be
deprecated in an upcoming release.
Added support for controlling whether a future is resolved
eagerly or lazily when creating the future,
e.g. future(..., lazy = TRUE),
futureAssign(..., lazy = TRUE), and
x %<-% { ... } %lazy% TRUE.
future(), futureAssign() and
futureCall() gained argument seed, which
specifies a L’Ecuyer-CMRG random seed to be used by the future. The seed
for future assignment can be specified via %seed%.
futureAssign() now passes all additional arguments
to future().
Added future_lapply() which supports load balancing
(“chunking”) and perfect reproducibility (regardless of type of load
balancing and how futures are resolved) via initial random
seed.
Added availableWorkers(). By default it returns
localhost workers according to availableCores(). In
addition, it detects common HPC allocations given in environment
variables set by the HPC scheduler.
The default for plan(cluster) is now
workers = availableWorkers().
Now plan() stops any clusters that were implicitly
created. For instance, a multisession cluster created by
plan(multisession) will be stopped when
plan(eager) is called.
makeClusterPSOCK() treats workers that refer to a
local machine by its local or canonical hostname as “localhost”. This
avoids having to launch such workers over SSH, which may not be
supported on all systems / compute cluster.
Option future.debug = TRUE also reports on total
size of globals identified and for cluster futures also the size of the
individual global variables exported.
Option future.wait.timeout (replaces
future.wait.times) specifies the maximum waiting time for a
free workers (e.g. a core or a compute node) before generating a timeout
error.
Option future.availableCores.fallback, which
defaults to environment variable
R_FUTURE_AVAILABLECORES_FALLBACK can now be used to specify
the default number of cores / workers returned by
availableCores() and availableWorkers() when
no other settings are available. For instance, if
R_FUTURE_AVAILABLECORES_FALLBACK=1 is set system wide in an
HPC environment, then all R processes that uses
availableCores() to detect how many cores can be used will
run as single-core processes. Without this fallback setting, and without
other core-specifying settings, the default will be to use all cores on
the machine, which does not play well on multi-user systems.
plan(lazy) are now deprecated.
Instead, use plan(eager) and then
f <- future(..., lazy = TRUE) or
x %<-% { ... } %lazy% TRUE. The reason behind this is
that in some cases code that uses futures only works under eager
evaluation (lazy = FALSE; the default), or vice verse. By
removing the “lazy” future strategy, the user can no longer override the
lazy = TRUE / FALSE that the developer is using.Creation of cluster futures (including multisession ones) would
time out already after 40 seconds if all workers were busy. New default
timeout is 30 days (option future.wait.timeout).
nbrOfWorkers() gave an error for
plan(cluster, workers) where workers was a
character vector or a cluster object of the
parallel package. Because of this,
future_lapply() gave an error with such setups.
availableCores(methods = "_R_CHECK_LIMIT_CORES_")
would give an error if not running R CMD check.
Added makeClusterPSOCK() - a version of
parallel::makePSOCKcluster() that allows for more flexible
control of how PSOCK cluster workers are set up and how they are
launched and communicated with if running on external machines.
Added generic as.cluster() for coercing objects to
cluster objects to be used as in
plan(cluster, workers = as.cluster(x)). Also added a
c() implementation for cluster objects such that multiple
cluster objects can be combined into a single one.
Added sessionDetails() for gathering details of the
current R session.
plan() and plan("list") now prints more
user-friendly output.
On Unix, internal myInternalIP() tries more
alternatives for finding the local IP number.
%<=% is deprecated. Use %<-%
instead. Same for %=>%.values() for lists and list environments of futures
where one or more of the futures resolved to NULL would give an
error.
value() for ClusterFuture would give
cryptic error message “Error in stop(ex) : bad error message” if the
cluster worker had crashed / terminated. Now it will instead give an
error message like “Failed to retrieve the value of
ClusterFuture from cluster node #1 on ‘localhost’. The
reason reported was”error reading from connection”.
Argument user to remote() was ignored
(since 1.1.0).
workers = "localhost" they (again) use the exact same R
executable as the main / calling R session (in all other cases it uses
whatever Rscript is found in the PATH). This
was already indeed implemented in 1.0.1, but with the added support for
reverse SSH tunnels in 1.1.0 this default behavior was lost.REMOTE CLUSTERS: It is now very simple to use
cluster() and remote() to connect to remote
clusters / machines. As long as you can connect via SSH to those
machines, it works also with these future. The new code completely
avoids incoming firewall and incoming port forwarding issues previously
needed. This is done by using reverse SSH tunneling. There is also no
need to worry about internal or external IP numbers.
Added optional argument label to all futures, e.g.
f <- future(42, label = "answer") and
v %<-% { 42 } %label% "answer".
Added argument user to cluster() and
remote().
Now all Future classes supports run()
for launching the future and value() calls
run() if the future has not been launched.
MEMORY: Now plan(cluster, gc = TRUE) causes the
background R session to be garbage collected immediately after the value
is collected. Since multisession and remote futures are special cases of
cluster futures, the same is true for these as well.
ROBUSTNESS: Now the default future strategy is explicitly set
when no strategies are set, e.g. when used nested futures. Previously,
only mc.cores was set so that only a single core was used, but now also
plan("default") set.
WORKAROUND: resolved() on cluster futures would
block on Linux until future was resolved. This is due to a bug in R. The
workaround is to use round the timeout (in seconds) to an integer, which
seems to always work / be respected.
Global variables part of subassignments in future expressions are
recognized and exported (iff found), e.g. x$a <- value,
x[["a"]] <- value, and
x[1,2,3] <- value.
Global variables part of formulae in future expressions are
recognized and exported (iff found),
e.g. y ~ x | z.
As an alternative to the default automatic identification of
globals, it is now also possible to explicitly specify them either by
their names (as a character vector) or by their names and values (as a
named list), e.g. f <- future({ 2*a }, globals = c("a"))
or f <- future({ 2*a }, globals = list(a = 42)). For
future assignments one can use the %globals% operator,
e.g. y %<-% { 2*a } %globals% c("a").
ROBUSTNESS: For the special case where ‘remote’ futures use
workers = "localhost" they now use the exact same R
executable as the main / calling R session (in all other cases it uses
whatever Rscript is found in the
PATH).
FutureError now extends simpleError and
no longer the error class of captured errors.
Since future 0.13.0, a global pkg
would be overwritten by the name of the last package attached in
future.
Futures that generated R.oo::Exception errors, they
triggered another internal error.
Add support for
remote(..., myip = "<external>"), which now queries a
set of external lookup services in case one of them fails.
Add mandelbrot() function used in demo to the API
for convenience.
ROBUSTNESS: If .future.R script, which is sourced
when the future package is attached, gives an error,
then the error is ignored with a warning.
TROUBLESHOOTING: If the future requires attachment of packages,
then each namespace is loaded separately and before attaching the
package. This is done in order to see the actual error message in case
there is a problem while loading the namespace. With
require()/library() this error message is
otherwise suppressed and replaced with a generic one.
Falsely identified global variables no longer generate an error
when the future is created. Instead, we leave it to R and the evaluation
of the individual futures to throw an error if the a global variable is
truly missing. This was done in order to automatically handle future
expressions that use non-standard evaluation (NSE),
e.g. subset(df, x < 3) where x is falsely
identified as a global variable.
Dropped support for system environment variable
R_FUTURE_GLOBALS_MAXSIZE.
DEMO: Now the Mandelbrot demo tiles a single Mandelbrot region with one future per tile. This better illustrates parallelism.
Documented R options used by the future package.
Custom futures based on a constructor function that is defined outside a package gave an error.
plan("default") assumed that the
future.plan option was a string; gave an error if it was a
function.
Various future options were not passed on to futures.
A startup .future.R script is no longer sourced if
the future package is attached by a future
expression.
Added remote futures, which are cluster futures with convenient
default arguments for simple remote access to R, e.g.
plan(remote, workers = "login.my-server.org").
Now .future.R (if found in the current directory or
otherwise in the user’s home directory) is sourced when the
future package is attach (but not loaded). This helps
separating scripts from configuration of futures.
Added support for
plan(cluster, workers = c("n1", "n2", "n2", "n4")), where
workers (also for ClusterFuture()) is a set of
host names passed to parallel::makeCluster(workers). It can
also be the number of localhost workers.
Added command line option --parallel=<p>,
which is long for -p <p>.
Now command line option -p <p> also set the
default future strategy to multiprocessing (if p >= 2 and eager
otherwise), unless another strategy is already specified via option
future.plan or system environment variable
R_FUTURE_PLAN.
Now availableCores() also acknowledges environment
variable NSLOTS set by Sun/Oracle Grid Engine
(SGE).
MEMORY: Added argument gc = FALSE to all futures.
When TRUE, the garbage collector will run at the very end in the process
that evaluated the future (just before returning the value). This may
help lowering the overall memory footprint when running multiple
parallel R processes. The user can enable this by specifying
plan(multiprocess, gc = TRUE). The developer can control
this using future(expr, gc = TRUE) or
v %<-% { expr } %tweak% list(gc = TRUE).
plan(list("eager")),
whereas it did work with plan("eager") and
plan(list(eager)).Added nbrOfWorkers().
Added informative print() method for the
Future class.
values() passes arguments ... to
value() of each future.
Added FutureError class.
maxCores and cluster to
workers. If using the old argument names a deprecation
warning will be generated, but it will still work until made defunct in
a future release.resolve() for lists and environments did not work
properly when the set of futures was not resolved in order, which could
happen with asynchronous futures.Add support to plan() for specifying different
future strategies for the different levels of nested futures.
Add backtrace() for listing the trace the
expressions evaluated (the calls made) before a condition was
caught.
Add transparent futures, which are eager futures with early signaling of conditioned enabled and whose expression is evaluated in the calling environment. This makes the evaluation of such futures as similar as possible to how R evaluates expressions, which in turn simplifies troubleshooting errors, etc.
Add support for early signaling of conditions. The default is (as before) to signal conditions when the value is queried. In addition, they may be signals as soon as possible, e.g. when checking whether a future is resolved or not.
Signaling of conditions when calling value() is now
controlled by argument signal (previously
onError).
Now UniprocessFuture:s captures the call stack for
errors occurring while resolving futures.
ClusterFuture() gained argument
persistent = FALSE. With persistent = TRUE,
any objects in the cluster R session that was created during the
evaluation of a previous future is available for succeeding futures that
are evaluated in the same session. Moreover, globals are still
identified and exported but “missing” globals will not give an error -
instead it is assumed such globals are available in the environment
where the future is evaluated.
OVERHEAD: Utility functions exported by
ClusterFuture are now much smaller; previously they would
export all of the package environment.
f <- multicore(NA, maxCores = 2) would end up in
an endless waiting loop for a free core if availableCores()
returned one.
ClusterFuture() would ignore
local = TRUE.
Added multiprocess futures, which are multicore futures if
supported, otherwise multisession futures. This makes it possible to use
plan(multiprocess) everywhere regardless of operating
system.
Future strategy functions gained class attributes such that it is
possible to test what type of future is currently used, e.g.
inherits(plan(), "multicore").
ROBUSTNESS: It is only the R process that created a future that can resolve it. If a non-resolved future is queried by another R process, then an informative error is generated explaining that this is not possible.
ROBUSTNESS: Now value() for multicore futures
detects if the underlying forked R process was terminated before
completing and if so generates an informative error messages.
resolve() gained argument
recursive.
Added option future.globals.resolve for controlling
whether global variables should be resolved for futures or not. If TRUE,
then globals are searched recursively for any futures and if found such
“global” futures are resolved. If FALSE, global futures are not located,
but if they are later trying to be resolved by the parent future, then
an informative error message is generated clarifying that only the R
process that created the future can resolve it. The default is currently
FALSE.
FIX: Exports of objects available in packages already attached by the future were still exported.
FIX: Now availableCores() returns 3L
(=2L+1L) instead of 2L if
_R_CHECK_LIMIT_CORES_ is set.
Add multisession futures, which analogously to multicore ones, use multiple cores on the local machine with the difference that they are evaluated in separate R session running in the background rather than separate forked R processes. A multisession future is a special type of cluster futures that do not require explicit setup of cluster nodes.
Add support for cluster futures, which can make use of a cluster
of nodes created by parallel::makeCluster().
Add futureCall(), which is for futures what
do.call() is otherwise.
Standardized how options are named,
i.e. future.<option>. If you used any future options
previously, make sure to check they follow the above format.
globals = TRUE).Now %<=% can also assign to multi-dimensional
list environments.
Add futures(), values() and
resolved().
Add resolve() to resolve futures in lists and
environments.
Now availableCores() also acknowledges the number of
CPUs allotted by Slurm.
CLEANUP: Now the internal future variable created by
%<=% is removed when the future variable is
resolved.
futureOf(envir = x) did not work properly when
x was a list environment.ROBUSTNESS: Now values of environment variables are trimmed before being parsed.
ROBUSTNESS: Add reproducibility test for random number generation using Pierre L’Ecuyer’s RNG stream regardless of how futures are evaluated, e.g. eager, lazy and multicore.
findGlobals(..., method = "ordered") in
globals (> 0.5.0) such that a global variable
preceding a local variable with the same name is properly identified and
exported/frozen.Globals that were copies of package objects were not exported to the future environments.
The future package had to be attached or
future::future() had to be imported, if
%<=% was used internally in another package. Similarly,
it also had to be attached if multicore futures where used.
eager() and multicore() gained argument
globals, where globals = TRUE will validate
that all global variables identified can be located already before the
future is created. This provides the means for providing the same tests
on global variables with eager and multicore futures as with lazy
futures.lazy(sum(x, ...), globals = TRUE) now properly passes
... from the function from which the future is setup. If
not called within a function or called within a function without
... arguments, an informative error message is thrown.plan("default") resets to the default strategy,
which is synchronous eager evaluation unless option
future_plan or environment variable
R_FUTURE_PLAN has been set.
availableCores("mc.cores") returns
getOption("mc.cores") + 1L, because option
mc.cores specifies “allowed number of additional R
processes” to be used in addition to the main R process.
plan(future::lazy) and similar gave errors.multicore() gained argument maxCores,
which makes it possible to use for instance
plan(multicore, maxCores = 4L).
Add availableMulticore() [from (in-house)
async package].
demo("mandelbrot", package = "future").ROBUSTNESS: multicore() blocks until one of the CPU
cores is available, iff all are currently occupied by other multicore
futures.
old <- plan(new) now returns the old
plan/strategy (was the newly set one).
Eager and lazy futures now records the result internally such that the expression is only evaluated once, even if their error values are requested multiple times.
Eager futures are always created regardless of error or not.
All Future objects are environments themselves that
record the expression, the call environment and optional
variables.
lazy() “freezes” global variables at the time when the
future is created. This way the result of a lazy future is more likely
to be the same as an ‘eager’ future. This is also how globals are likely
to be handled by asynchronous futures.plan() records the call.demo("mandelbrot", package = "future"), which can
be re-used by other future packages.Added plan().
Added eager future - useful for troubleshooting.